AI 改寫工具與檢測器 2026 實測：Phrasly 令 3 款檢測器判為 0% AI，但刪去原文細節

By Marcus Hale.

Search “AI humanizer” and every result promises the same thing: paste your ChatGPT essay, click a button, walk away with text that sails past Turnitin at a 99% success rate. The pages selling these tools are confident. The reviews ranking them are usually written by the same people selling them.

So I ran the test myself. One paragraph of raw ChatGPT output, four detectors, three humanizers, and a simple question: does any of this actually work? The answer turned out to be more useful than either side of the marketing war will tell you, and it ends somewhere most “best humanizer” lists never go: a New York courtroom and a list of universities that have quietly switched their detectors off.

What an AI humanizer actually does

An AI humanizer rewrites machine-generated text so it reads less like a machine wrote it. That’s the whole pitch. The interesting part is what “less like a machine” means in math.

Detectors lean on two measurements. The first is perplexity, which tracks how predictable each word is. Language models are built to pick the most probable next word, so their output runs unusually smooth and predictable. People don’t write that way. We reach for the odd word, the unexpected turn, the phrasing nobody saw coming. The second is burstiness, the variation in sentence length and rhythm. Human writing swings from four-word sentences to sprawling ones; AI tends to settle into an even, uniform cadence.

Humanizers exist to disrupt both signals. The cheap ones swap words for synonyms. The better ones restructure whole sentences, vary the pacing, and break the parallel-clause habit that gives models away. The most aggressive ones rewrite the passage from scratch under a different statistical fingerprint. More on what that costs in a minute.

Perplexity = how predictable the words are (low = robotic). Burstiness = how much sentence length varies (low = robotic). Every detector and every humanizer is fighting over these two numbers.

The test: raw AI text against four detectors

I started with a clean control: a single paragraph generated by ChatGPT about AI in business, 98 words, untouched. Then I ran it through four free detectors anyone can use without paying: GPTZero, Scribbr, QuillBot, and ZeroGPT.

They didn’t hesitate. GPTZero, Scribbr, and QuillBot all returned 100% AI. ZeroGPT came in at 92.7%. Three out of four called it machine-written with full confidence, and the fourth wasn’t far behind.

These are the tools’ scores for one sample, not measured accuracy rates.

That’s worth sitting with before we go further: on this untouched sample, all four detectors flagged the AI text. The marketing for humanizers depends on you forgetting that the starting point is a 100% flag.

Then I ran it through humanizers. Here’s the catch

I put the same paragraph through three humanizers: QuillBot’s humanizer on its standard setting, HumanizeAI.pro on free mode, and Phrasly on its most aggressive setting. Then I sent each rewrite back through the detectors.

The two lighter tools failed. QuillBot’s humanized version still scored 100% AI, and it failed against QuillBot’s own detector, which is the same company grading its own homework and giving itself an F. HumanizeAI.pro’s version came back 100% AI on GPTZero, which added a second flag: “Possible AI Paraphrasing,” with a verdict reading “originally AI, but rewritten by AI or a human.” The detector didn’t just catch the text. It caught the act of running it through a humanizer.

Then there was Phrasly on aggressive mode. That one worked. GPTZero, Scribbr, and QuillBot all flipped to 0% AI, “entirely human,” in GPTZero’s words. Only ZeroGPT held the line at 62.8%. Three detectors out of four, beaten.

Here’s the catch the tool sellers leave out. I read the version that beat the detectors. To get there, Phrasly had gutted it. The original paragraph mentioned reducing operational costs, analyzing vast amounts of data, maintaining a competitive edge, and freeing teams for strategic initiatives. The “human” version that passed dropped all of it. What survived was a string of short, flat sentences (“Companies are using AI to better themselves and make work more efficient”) that say almost nothing. It read as human because it read as generic. The specificity, the substance, the reason anyone would write the paragraph in the first place: gone.

Meikuio test, June 2026. The rewrite also removed specific meaning from the paragraph.

The choice a humanizer really gives you is undetectable or good. Light edits keep your meaning but get caught. Aggressive edits slip past most detectors by flattening the writing into something blander than what you started with.

What the detectors’ own makers admit

The vendors on both sides quote accuracy numbers that don’t survive contact with their own fine print.

Turnitin advertises 98% accuracy with a false-positive rate under 1%. But that figure only applies to documents already scored at 20% AI or higher, and Turnitin’s own Chief Product Officer, Annie Chechitelli, has said the real detection rate is lower: “We estimate that we find about 85% of AI writing. We let probably 15% go by in order to reduce our false positives to less than 1 percent.” So the headline number is 98%, the operational reality is closer to 85%, and even the sub-1% false-positive claim is a per-document figure. At the sentence level, Turnitin’s own documentation puts the false-positive rate around 4%.

GPTZero tells a similar story from the other direction. Its own 2026 benchmark claims 99.4% accuracy and a false-positive rate of 0.24%, roughly one wrong flag in 400. Independent testing in early 2026 found something rougher: around 87% accuracy with a 10% false-positive rate in one benchmark, and false-positive rates ranging anywhere from 2% to 29% depending on the test set. Both numbers come from real tests. The gap between them is the whole problem with detection benchmarks: score a tool on the kind of text it was tuned for and it looks flawless, change the text and it doesn’t.

The most honest data point in this whole field comes from OpenAI. In January 2023 it released its own AI text classifier. It shut the tool down six months later, in July 2023, citing a “low rate of accuracy.” The numbers were brutal: it correctly identified just 26% of AI-written text and falsely flagged 9% of human writing. The company that built ChatGPT, with full access to the model, could not reliably detect the model’s own output. That should tell you how solid the ground is under everyone else’s 99% claims.

Watch the upsell. When QuillBot and ZeroGPT flagged my text as AI, both showed a button right under the result (“Refine with Paraphraser,” “Humanize Text”) offering to hide what they’d just detected. The same companies profit from catching AI and from hiding it.

The part nobody selling a humanizer mentions: false positives

Everything above is about catching AI. The bigger danger runs the other way: detectors flagging real human writing as machine-made. It’s the same reliability problem we hit when we tested free deepfake detectors on images and video.

The landmark study here came out of Stanford in 2023, published in the journal Patterns by a team led by Weixin Liang under Professor James Zou. They ran 91 TOEFL essays, written by real non-native English speakers, through seven popular AI detectors. The detectors flagged 61% of those genuine human essays as AI-generated on average. Nearly one in five was flagged unanimously by all seven tools, and 97.8% were flagged by at least one. The mechanism is bias baked into the math: non-native writers use a narrower vocabulary and simpler sentence structures, which look statistically identical to the low-perplexity patterns detectors are trained to catch.

It isn’t only non-native speakers. Neurodivergent students, who often write in consistent, repetitive patterns, trip the same wires. So does anyone who runs their work through Grammarly, which smooths out the natural variation in writing and pushes it closer to the machine zone. And the editing fragility is just as bad. A 2024 study by Perkins and colleagues found that six detectors averaging a modest 39.5% accuracy on AI text dropped to 17.4% the moment that text was lightly edited to evade them. Newer research from 2026 pushed this to an extreme: a reinforcement-learning paraphraser called StealthRL drove detection down to near zero with a 99.9% evasion rate against a panel of detectors.

Turnitin has fought back. In August 2025 it rolled out “AI bypasser detection,” built specifically to flag text that’s been run through paraphrasing and humanizing tools, which is exactly what caught my HumanizeAI sample on GPTZero. But that arms-race patch came with its own reported 6% false-positive rate on genuine human writing. Every turn of the screw to catch the cheaters catches more innocent students too.

Universities are switching detectors off

Faced with those numbers, a growing list of institutions have stopped trusting the tools. Vanderbilt University disabled Turnitin’s AI detector back in 2023, after students using Grammarly got wrongly flagged. The University of Waterloo followed in September 2025, citing unreliability and the bias against non-native speakers, noting its own testing flagged human-written text as 100% AI.

Washington State University cancelled its Turnitin AI-detection contract in February 2026, and its reasoning is the clearest case study in the field. In a single semester, Fall 2024, the tool scanned 148,547 assessments. Even at a 1% false-positive rate, that’s roughly 1,485 students wrongly flagged. And between 2023 and 2025, 33% of the school’s AI-misconduct cases ended in a “not responsible” finding because the detection score was the only evidence anyone had. Australia’s Curtin University switched its detector off in January 2026, and Australian Catholic University abandoned its tool after logging nearly 6,000 misconduct allegations in 2024, about 90% of them AI-related, with a large share dismissed on review.

In one semester Washington State ran 148,547 assessments through Turnitin. A third of the AI cases that reached its review board ended in acquittal, because a detector score was the only proof.

A court just threw out an AI-cheating accusation

The clearest signal of where this is heading came from a courtroom. In Newby v. Adelphi University, decided in Nassau County Supreme Court on January 28, 2026, the court annulled a university’s AI-cheating finding outright.

The student, Orion Newby, was an autistic freshman enrolled in Adelphi’s Bridges support program for neurodivergent students. His World Civilizations essay was run through Turnitin, which flagged it at 100% AI. The university handed him a zero and an academic-integrity violation. Newby pushed back with evidence: independent checks from Grammarly and ZeroGPT both read his essay as human-written, and Turnitin’s own originality report showed just 4% overlap with existing sources. The court ruled the university’s decision “without valid basis and devoid of reason” and ordered the violation expunged from his record. A former U.S. Attorney called the opinion “groundbreaking.”

The ruling is narrow and case-specific. The court set aside the violation after finding that the school had ignored contrary evidence, denied the student a meaningful appeal, and relied on a report the professor later undermined. It did not establish a blanket rule for every detector score. A score alone deserves scrutiny, and the surrounding evidence and process matter.

So should you use one?

Strip away both marketing campaigns and here’s what my test and the research actually support.

If you’re writing with AI and hoping a humanizer makes it undetectable, the honest read is that you’re picking between two bad outcomes. A light pass keeps your meaning but gets caught, and now gets flagged as paraphrased on top of it. An aggressive pass can slip past most detectors, but it strips your work down to bland filler, and even then one detector out of four still caught it. There’s no setting that gives you good writing and reliable invisibility at once.

If you’re a student reaching for a humanizer defensively, because you wrote the thing yourself and you’re scared a detector will falsely flag you, that fear is rational, given a 61% false-positive rate on non-native essays. But running clean writing through a humanizer can backfire, since the rewrite itself now trips bypasser detection. The better protection is a record of your process: keep your drafts, your version history, your notes. That’s also where the smart institutions are heading, away from after-the-fact detection and toward asking students to show their work. The Newby court set aside a disciplinary finding after the university ignored contrary evidence and mishandled the appeal. Your draft history can help show how the work developed.

Falsely accused of using AI? Don’t panic and don’t reword anything. Gather your drafts and version history (Google Docs and Word both timestamp edits), get a second opinion from another detector, and ask the school for the actual report and its appeal process. The Newby decision shows why the school must examine the evidence and follow its own appeal rules. (The same “verify, don’t trust the tool” instinct protects you from AI voice scams too.)

Frequently asked questions

Can Turnitin detect AI humanizers in 2026?

Often, yes. In August 2025 Turnitin added “AI bypasser detection” aimed specifically at text run through paraphrasing and humanizing tools, and it now flags suspected paraphrased AI as a separate category. In my own test, GPTZero similarly flagged a humanized paragraph as “Possible AI Paraphrasing.” Light humanizing tends to get caught, and the rewrite itself can become a red flag.

Do AI humanizers really work?

It depends entirely on how aggressively they rewrite. In my test, light and medium humanizers left the text at 100% AI. Only an aggressive setting beat three of four detectors, and it did so by flattening the writing into generic, low-information text. The “99% bypass” claims vendors advertise don’t hold up across detectors or without wrecking quality.

Will a humanizer get me caught?

Possibly more than the original AI text would. Several detectors now flag the signature of bypasser tools, so a humanized passage can be marked as paraphrased even when a plain AI passage might have scraped by. There’s no guarantee of safety in either direction.

Are AI detectors accurate?

They’re good at catching untouched AI text and unreliable on edited text or human writing. Turnitin’s own product chief puts real detection near 85%, not the advertised 98%. Independent tests show false-positive rates from a few percent up to 29%, and a Stanford study found 61% of genuine non-native English essays were wrongly flagged. Accurate enough to raise a question, not to prove one.

Is using an AI humanizer cheating?

If you’re using it to disguise AI-written work as your own to get around your school’s rules, most academic-integrity policies would treat that as a violation, regardless of whether the text gets detected. If you’re an ESL writer or neurodivergent student smoothing your own genuine writing, that’s a grayer area, but be aware the rewrite can trip bypasser detection and create the very suspicion you were trying to avoid.

What should I do if I’m falsely accused of using AI?

Keep your drafts and version history, which timestamp your writing process and are the strongest evidence you have. Get a second reading from another detector, request the actual report from your instructor, and follow the formal appeal process. In Newby v. Adelphi, the court set aside a finding after the university failed to consider contrary evidence and provide a meaningful appeal. The ruling turned on that record and process.

Originally published on Meikuio — AI Humanizers vs. Detectors in 2026: I Tested Them. Here's What Actually Happens.